1 %************************************************************************
3 \section[Driver-asm-fiddling]{Fiddling with assembler files}
5 %************************************************************************
10 Utterly stomp out C functions' prologues and epilogues; i.e., the
11 stuff to do with the C stack.
13 Any other required tidying up.
16 General note [chak]: Many regexps are very fragile because they rely on white
17 space being in the right place. This caused trouble with gcc 2.95 (at least
18 on Linux), where the use of white space in .s files generated by gcc suddenly
19 changed. To guarantee compatibility across different versions of gcc, make
20 sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21 space between an assembler statement and its arguments as well as after a the
22 comma separating multiple arguments.
24 \emph{For the time being, I have corrected the regexps for i386-.*-linux. I
25 didn't touch all the regexps for other i386 platforms, as I don't have
26 a box to test these changes.}
31 The HP linker is very picky about symbols being in the appropriate
32 space (code vs. data). When we mangle the threaded code to put the
33 info tables just prior to the code, they wind up in code space
34 rather than data space. This means that references to *_info from
35 un-mangled parts of the RTS (e.g. unthreaded GC code) get
36 unresolved symbols. Solution: mini-mangler for .c files on HP. I
37 think this should really be triggered in the driver by a new -rts
38 option, so that user code doesn't get mangled inappropriately.
40 With reversed tables, jumps are to the _info label rather than to
41 the _entry label. The _info label is just an address in code
42 space, rather than an entry point with the descriptive blob we
43 talked about yesterday. As a result, you can't use the call-style
44 JMP_ macro. However, some JMP_ macros take _info labels as targets
45 and some take code entry points within the RTS. The latter won't
46 work with the goto-style JMP_ macro. Sigh. Solution: Use the goto
47 style JMP_ macro, and mangle some more assembly, changing all
48 "RP'literal" and "LP'literal" references to "R'literal" and
49 "L'literal," so that you get the real address of the code, rather
50 than the descriptive blob. Also change all ".word P%literal"
51 entries in info tables and vector tables to just ".word literal,"
52 for the same reason. Advantage: No more ridiculous call sequences.
55 %************************************************************************
57 \subsection{Top-level code}
59 %************************************************************************
62 ############################################################################
63 # Make all regexp matching multi-line aware. This replaces the line below
64 # originally found in "sub mangle_asm":
68 # This used to work, but Perl 5.10 removes support for $*, so we uses an
69 # equivalent construct that works in Perl 5.6 and later.
71 BEGIN { require overload; overload::constant( qr => sub { "(?m:$_[1])" } ) }
72 ############################################################################
74 $TargetPlatform = $TARGETPLATFORM;
76 ($Pgm = $0) =~ s|.*/||;
80 if ( $TargetPlatform =~ /^i386-/ ) {
84 $StolenX86Regs = $ARGV[2];
88 &mangle_asm($ifile,$ofile);
93 %************************************************************************
95 \subsection{Constants for various architectures}
97 %************************************************************************
100 sub init_TARGET_STUFF {
102 #--------------------------------------------------------#
103 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
105 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
106 $T_US = ''; # _ if symbols have an underscore on the front
107 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
108 $T_CONST_LBL = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
111 $T_MOVE_DIRVS = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
112 $T_COPY_DIRVS = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
114 $T_DOT_WORD = '\.(long|quad|byte|word)';
115 $T_DOT_GLOBAL = '^\t\.globl';
116 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
117 $T_HDR_misc = "\.text\n\t\.align 3\n";
118 $T_HDR_data = "\.data\n\t\.align 3\n";
119 $T_HDR_rodata = "\.rdata\n\t\.align 3\n";
120 $T_HDR_closure = "\.data\n\t\.align 3\n";
121 $T_HDR_info = "\.text\n\t\.align 3\n";
122 $T_HDR_entry = "\.text\n\t\.align 3\n";
123 $T_HDR_vector = "\.text\n\t\.align 3\n";
125 #--------------------------------------------------------#
126 } elsif ( $TargetPlatform =~ /^hppa/ ) {
128 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
129 $T_US = ''; # _ if symbols have an underscore on the front
130 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
131 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
134 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
135 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
137 $T_DOT_WORD = '\.(blockz|word|half|byte)';
138 $T_DOT_GLOBAL = '^\s+\.EXPORT';
139 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
140 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
141 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
142 $T_HDR_rodata = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
143 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
144 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
145 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
146 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
148 #--------------------------------------------------------#
149 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
150 # NeXT added but not tested. CaS
152 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
153 $T_US = '_'; # _ if symbols have an underscore on the front
154 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
155 $T_CONST_LBL = '^LC(\d+):$';
157 $T_X86_PRE_LLBL_PAT = 'L';
158 $T_X86_PRE_LLBL = 'L';
159 $T_X86_BADJMP = '^\tjmp [^L\*]';
161 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
162 $T_COPY_DIRVS = '\.(globl|stab|lcomm)';
163 $T_DOT_WORD = '\.(long|word|value|byte|space)';
164 $T_DOT_GLOBAL = '\.globl';
165 $T_HDR_literal = "\.text\n\t\.align 2\n";
166 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
167 $T_HDR_data = "\.data\n\t\.align 2\n";
168 $T_HDR_rodata = "\.text\n\t\.align 2\n";
169 $T_HDR_closure = "\.data\n\t\.align 2\n";
170 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
171 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
172 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
174 #--------------------------------------------------------#
175 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
177 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
178 $T_US = ''; # _ if symbols have an underscore on the front
179 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
180 ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
181 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
183 $T_X86_PRE_LLBL_PAT = '\.L';
184 $T_X86_PRE_LLBL = '.L';
185 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
187 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
188 if ( $TargetPlatform =~ /solaris2/ ) {
189 # newer Solaris linkers are picky about .size information, so
190 # omit it (see #1421)
191 $T_COPY_DIRVS = '^\s*\.(globl|local)';
193 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
196 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
197 $T_DOT_GLOBAL = '\.globl';
198 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
199 $T_HDR_misc = "\.text\n\t\.align 4\n";
200 $T_HDR_data = "\.data\n\t\.align 4\n";
201 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
202 $T_HDR_closure = "\.data\n\t\.align 4\n";
203 $T_HDR_info = "\.text\n\t\.align 4\n";
204 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
205 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
207 #--------------------------------------------------------#
208 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
210 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
211 $T_US = ''; # _ if symbols have an underscore on the front
213 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
216 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
217 $T_COPY_DIRVS = '\.(global|proc)';
219 $T_DOT_WORD = '\.(long|value|byte|zero)';
220 $T_DOT_GLOBAL = '\.global';
221 $T_HDR_literal = "\.section\t\.rodata\n";
222 $T_HDR_misc = "\.text\n\t\.align 16\n"; # May contain code; align like 'entry'
223 $T_HDR_data = "\.data\n\t\.align 8\n";
224 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
225 $T_HDR_closure = "\.data\n\t\.align 8\n";
226 $T_HDR_info = "\.text\n\t\.align 8\n";
227 $T_HDR_entry = "\.text\n\t\.align 16\n";
228 $T_HDR_vector = "\.text\n\t\.align 8\n";
230 #--------------------------------------------------------#
231 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
233 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
234 $T_US = ''; # _ if symbols have an underscore on the front
236 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
239 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)([ \t].*)?\n)';
240 $T_COPY_DIRVS = '\.(globl|type|size|local)';
242 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
243 $T_DOT_GLOBAL = '\.global';
245 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
246 $T_HDR_literal = "\.section\t\.rodata\n";
248 $T_HDR_misc = "\.text\n\t\.align 8\n";
249 $T_HDR_data = "\.data\n\t\.align 8\n";
250 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
252 # the assembler on x86_64/Linux refuses to generate code for
254 # where x is in the text section and y in the rodata section.
255 # It works if y is in the text section, though. This is probably
256 # going to cause difficulties for PIC, I imagine.
258 # See Note [x86-64-relative] in includes/InfoTables.h
259 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
261 $T_HDR_closure = "\.data\n\t\.align 8\n";
262 $T_HDR_info = "\.text\n\t\.align 8\n";
263 $T_HDR_entry = "\.text\n\t\.align 8\n";
264 $T_HDR_vector = "\.text\n\t\.align 8\n";
266 #--------------------------------------------------------#
267 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
269 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
270 $T_US = '_'; # _ if symbols have an underscore on the front
271 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
272 $T_CONST_LBL = '^LC(\d+):$';
275 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
276 $T_COPY_DIRVS = '\.(globl|proc|stab)';
278 $T_DOT_WORD = '\.long';
279 $T_DOT_GLOBAL = '\.globl';
280 $T_HDR_literal = "\.text\n\t\.even\n";
281 $T_HDR_misc = "\.text\n\t\.even\n";
282 $T_HDR_data = "\.data\n\t\.even\n";
283 $T_HDR_rodata = "\.text\n\t\.even\n";
284 $T_HDR_closure = "\.data\n\t\.even\n";
285 $T_HDR_info = "\.text\n\t\.even\n";
286 $T_HDR_entry = "\.text\n\t\.even\n";
287 $T_HDR_vector = "\.text\n\t\.even\n";
289 #--------------------------------------------------------#
290 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
292 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
293 $T_US = ''; # _ if symbols have an underscore on the front
294 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
295 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
298 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
299 $T_COPY_DIRVS = '\.(globl|ent)';
301 $T_DOT_WORD = '\.word';
302 $T_DOT_GLOBAL = '^\t\.globl';
303 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
304 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
305 $T_HDR_data = "\t\.data\n\t\.align 2\n";
306 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
307 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
308 $T_HDR_info = "\t\.text\n\t\.align 2\n";
309 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
310 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
312 #--------------------------------------------------------#
313 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
314 # Apple PowerPC Darwin/MacOS X.
315 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
316 $T_US = '_'; # _ if symbols have an underscore on the front
317 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
318 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
321 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
322 $T_COPY_DIRVS = '\.(globl|lcomm)';
324 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
325 $T_DOT_GLOBAL = '\.globl';
326 $T_HDR_toc = "\.toc\n";
327 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
328 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
329 $T_HDR_data = "\t\.data\n\t\.align 2\n";
330 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
331 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
332 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
333 $T_HDR_info = "\t\.text\n\t\.align 2\n";
334 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
335 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
337 #--------------------------------------------------------#
338 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
339 # Apple PowerPC Darwin/MacOS X.
340 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
341 $T_US = '_'; # _ if symbols have an underscore on the front
342 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
343 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
345 $T_X86_PRE_LLBL_PAT = 'L';
346 $T_X86_PRE_LLBL = 'L';
347 $T_X86_BADJMP = '^\tjmp [^L\*]';
349 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
350 $T_COPY_DIRVS = '\.(globl|lcomm)';
352 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
353 $T_DOT_GLOBAL = '\.globl';
354 $T_HDR_toc = "\.toc\n";
355 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
356 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
357 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
358 $T_HDR_data = "\t\.data\n\t\.align 2\n";
359 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
360 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
361 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
362 $T_HDR_info = "\t\.text\n\t\.align 2\n";
363 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
364 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
366 #--------------------------------------------------------#
367 } elsif ( $TargetPlatform =~ /^x86_64-apple-darwin.*/ ) {
368 # Apple PowerPC Darwin/MacOS X.
369 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
370 $T_US = '_'; # _ if symbols have an underscore on the front
371 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
372 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
375 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
376 $T_COPY_DIRVS = '\.(globl|lcomm)';
378 $T_DOT_WORD = '\.(quad|long|short|byte|fill|space)';
379 $T_DOT_GLOBAL = '\.globl';
380 $T_HDR_toc = "\.toc\n";
381 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
382 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
383 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
384 $T_HDR_data = "\t\.data\n\t\.align 2\n";
385 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
386 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
387 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
388 $T_HDR_info = "\t\.text\n\t\.align 2\n";
389 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
390 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
392 #--------------------------------------------------------#
393 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
395 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
396 $T_US = ''; # _ if symbols have an underscore on the front
397 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
398 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
401 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
402 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
404 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
405 $T_DOT_GLOBAL = '\.globl';
406 $T_HDR_toc = "\.toc\n";
407 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
408 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
409 $T_HDR_data = "\t\.data\n\t\.align 2\n";
410 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
411 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
412 $T_HDR_info = "\t\.text\n\t\.align 2\n";
413 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
414 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
416 #--------------------------------------------------------#
417 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
419 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
420 $T_US = '\.'; # _ if symbols have an underscore on the front
421 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
422 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
425 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
426 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
428 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
429 $T_DOT_GLOBAL = '\.globl';
430 $T_HDR_toc = "\.toc\n";
431 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
432 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
433 $T_HDR_data = "\t\.data\n\t\.align 2\n";
434 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
435 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
436 $T_HDR_info = "\t\.text\n\t\.align 2\n";
437 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
438 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
440 #--------------------------------------------------------#
441 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
443 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
444 $T_US = ''; # _ if symbols have an underscore on the front
445 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
446 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
449 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.stab.*|\s*\.section.*|\s+\.type.*|\s+\.size.*)\n)';
450 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
452 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
453 $T_DOT_GLOBAL = '^\t\.global';
454 $T_HDR_literal = "\.text\n\t\.align 8\n";
455 $T_HDR_misc = "\.text\n\t\.align 4\n";
456 $T_HDR_data = "\.data\n\t\.align 8\n";
457 $T_HDR_rodata = "\.text\n\t\.align 4\n";
458 $T_HDR_closure = "\.data\n\t\.align 4\n";
459 $T_HDR_info = "\.text\n\t\.align 4\n";
460 $T_HDR_entry = "\.text\n\t\.align 4\n";
461 $T_HDR_vector = "\.text\n\t\.align 4\n";
463 #--------------------------------------------------------#
464 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
466 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
467 $T_US = '_'; # _ if symbols have an underscore on the front
468 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
469 $T_CONST_LBL = '^LC(\d+):$';
472 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
473 $T_COPY_DIRVS = '\.(global|proc|stab)';
475 $T_DOT_WORD = '\.word';
476 $T_DOT_GLOBAL = '^\t\.global';
477 $T_HDR_literal = "\.text\n\t\.align 8\n";
478 $T_HDR_misc = "\.text\n\t\.align 4\n";
479 $T_HDR_data = "\.data\n\t\.align 8\n";
480 $T_HDR_rodata = "\.text\n\t\.align 4\n";
481 $T_HDR_closure = "\.data\n\t\.align 4\n";
482 $T_HDR_info = "\.text\n\t\.align 4\n";
483 $T_HDR_entry = "\.text\n\t\.align 4\n";
484 $T_HDR_vector = "\.text\n\t\.align 4\n";
486 #--------------------------------------------------------#
487 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
488 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
489 $T_US = ''; # _ if symbols have an underscore on the front
490 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
491 # Probably doesn't apply anyway
492 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
495 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.seg|\.stab.*|\s+?\.section.*|\s+\.type.*|\s+\.size.*)\n)';
496 $T_COPY_DIRVS = '\.(global|local|globl|proc|stab)';
498 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
499 $T_DOT_GLOBAL = '^\t\.global';
500 $T_HDR_literal = "\.text\n\t\.align 8\n";
501 $T_HDR_misc = "\.text\n\t\.align 4\n";
502 $T_HDR_data = "\.data\n\t\.align 8\n";
503 $T_HDR_rodata = "\.text\n\t\.align 4\n";
504 $T_HDR_closure = "\.data\n\t\.align 4\n";
505 $T_HDR_info = "\.text\n\t\.align 4\n";
506 $T_HDR_entry = "\.text\n\t\.align 4\n";
507 $T_HDR_vector = "\.text\n\t\.align 4\n";
509 #--------------------------------------------------------#
511 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
515 if($T_HDR_relrodata eq "") {
517 # relrodata defaults to rodata.
518 $T_HDR_relrodata = $T_HDR_rodata;
522 print STDERR "T_STABBY: $T_STABBY\n";
523 print STDERR "T_US: $T_US\n";
524 print STDERR "T_PRE_APP: $T_PRE_APP\n";
525 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
526 print STDERR "T_POST_LBL: $T_POST_LBL\n";
527 if ( $TargetPlatform =~ /^i386-/ ) {
528 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
529 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
530 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
532 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
533 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
534 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
535 print STDERR "T_HDR_literal: $T_HDR_literal\n";
536 print STDERR "T_HDR_misc: $T_HDR_misc\n";
537 print STDERR "T_HDR_data: $T_HDR_data\n";
538 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
539 print STDERR "T_HDR_closure: $T_HDR_closure\n";
540 print STDERR "T_HDR_info: $T_HDR_info\n";
541 print STDERR "T_HDR_entry: $T_HDR_entry\n";
542 print STDERR "T_HDR_vector: $T_HDR_vector\n";
548 %************************************************************************
550 \subsection{Mangle away}
552 %************************************************************************
556 local($in_asmf, $out_asmf) = @_;
559 # ia64-specific information for code chunks
563 &init_TARGET_STUFF();
564 &init_FUNNY_THINGS();
566 open(INASM, "< $in_asmf")
567 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
568 open(OUTASM,"> $out_asmf")
569 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
571 # read whole file, divide into "chunks":
572 # record some info about what we've found...
574 @chk = (); # contents of the chunk
575 $numchks = 0; # number of them
576 @chkcat = (); # what category of thing in each chunk
577 @chksymb = (); # what symbol(base) is defined in this chunk
578 %entrychk = (); # ditto, its entry code
579 %closurechk = (); # ditto, the (static) closure
580 %srtchk = (); # ditto, its SRT (for top-level things)
581 %infochk = (); # given a symbol base, say what chunk its info tbl is in
582 %vectorchk = (); # ditto, return vector table
583 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
585 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
588 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
589 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
590 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
591 next if /^\t\.def.*endef$/;
592 next if /${T_PRE_APP}(NO_)?APP/o;
593 next if /^;/ && $TargetPlatform =~ /^hppa/;
595 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
597 if ( $TargetPlatform =~ /^mips-/
598 && /^\t\.(globl\S+\.text|comm\t)/ ) {
599 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
600 # Treat .comm variables as data. These show up in two (known) places:
602 # - the module_registered variable used in the __stginit fragment.
603 # even though these are declared static and initialised, gcc 3.3
604 # likes to make them .comm, presumably to save space in the
607 # - global variables used to pass arguments from C to STG in
608 # a foreign export. (is this still true? --SDM)
610 } elsif ( /^\t\.comm.*$/ ) {
612 $chkcat[$i] = 'data';
615 # Labels ending "_str": these are literal strings.
616 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
618 $chkcat[$i] = 'relrodata';
620 } elsif ( $TargetPlatform =~ /-darwin/
621 && (/^\s*\.subsections_via_symbols/
622 ||/^\s*\.no_dead_strip.*/)) {
623 # Don't allow Apple's linker to do any dead-stripping of symbols
624 # in this file, because it will mess up info-tables in mangled
626 # The .no_dead_strip directives are actually put there by
627 # the gcc3 "used" attribute on entry points.
629 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
630 /^\s*\.picsymbol_stub/
631 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
632 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
633 || /^\s*\.symbol_stub/
634 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
635 || /^\s*\.section __TEXT,__symbolstub\d,.*/
636 || /^\s*\.lazy_symbol_pointer/
637 || /^\s*\.non_lazy_symbol_pointer/
638 || /^\s*\.section __IMPORT.*/))
641 $chkcat[$i] = 'dyld';
645 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
646 { # non_lazy_symbol_ptrs that point to local symbols
648 $chkcat[$i] = 'dyld';
651 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.align/)
652 { # non_lazy_symbol_ptrs that point to local symbols
654 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
655 { # non_lazy_symbol_ptrs that point to local symbols
656 $chk[++$i] = $dyld_section . $_;
657 $chkcat[$i] = 'dyld';
660 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
661 # duplicated from the bottom
665 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
666 # Alphas: Local labels not to be confused with new chunks
668 # NB: all the rest start with a non-space
670 } elsif ( $TargetPlatform =~ /^mips-/
671 && /^\d+:/ ) { # a funny-looking very-local label
674 } elsif ( /$T_CONST_LBL/o ) {
676 $chkcat[$i] = 'literal';
679 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
681 $chkcat[$i] = 'splitmarker';
684 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
687 $chkcat[$i] = 'infotbl';
688 $chksymb[$i] = $symb;
690 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
692 $infochk{$symb} = $i;
694 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
696 $chkcat[$i] = 'entry';
701 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
703 $chkcat[$i] = 'closure';
706 $closurechk{$1} = $i;
708 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
715 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
717 $chkcat[$i] = 'data';
720 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
722 $chkcat[$i] = 'data';
725 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
728 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
729 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
730 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
731 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
734 $chkcat[$i] = 'data';
737 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
742 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
743 # all CC_ symbols go in the data section...
745 $chkcat[$i] = 'data';
748 } elsif ( /^${T_US}([A-Za-z0-9_]+)_hpc${T_POST_LBL}$/o ) {
749 # hpc shares tick boxes across modules
751 $chkcat[$i] = 'data';
754 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
756 $chkcat[$i] = 'misc';
758 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
760 $chkcat[$i] = 'vector';
765 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
766 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
767 # Some Solaris system headers contain function definitions (as
768 # opposed to mere prototypes), which end up in the .hc file when
769 # a Haskell module foreign imports the corresponding system
770 # functions (most notably stat()). We put them into the text
771 # segment. Note that this currently does not extend to function
772 # names starting with an underscore.
775 $chkcat[$i] = 'misc';
778 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
779 # To handle PIC on Darwin/x86, we need to appropriately pass through
780 # the get_pc_thunk functions. The need to be put into a special section
781 # marked as coalesced (otherwise the .weak_definition doesn't work
784 $chkcat[$i] = 'get_pc_thunk';
787 } elsif ( /^${T_US}[A-Za-z0-9_]/o
788 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
790 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
791 || ! /^\.L\d+:$/ ) ) {
798 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
799 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
800 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
801 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
802 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
803 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
804 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
805 || /^_uname:/o # x86/Solaris2
808 $chkcat[$i] = 'misc';
810 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
811 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
814 $chkcat[$i] = 'relrodata';
817 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
818 $chkcat[$i] = 'unknown';
821 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
822 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
823 # table "by hand". Be sure to copy it over.
824 # Note that this label and all entries in the table should actually
825 # go into the .got2 section, but it isn't easy to distinguish them
826 # from other constant literals (.LC\d+), so we just put everything
829 $chkcat[$i] = 'literal';
830 $chksymb[$i] = 'LCTOC1';
831 } else { # simple line (duplicated at the top)
836 $numchks = $#chk + 1;
837 $chk[$numchks] = ''; # We might push .note.GNU-stack into this
838 $chkcat[$numchks] = 'verbatim'; # If we do, write it straight back out
840 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
841 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
844 # the division into chunks is imperfect;
845 # we throw some things over the fence into the next
848 # also, there are things we would like to know
849 # about the whole module before we start spitting
852 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
853 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
855 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
857 # Alphas: NB: we start meddling at chunk 1, not chunk 0
858 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
859 # spits a ".quad 0" in after the very first ".rdata"; we
860 # detect this special case (tossing the ".quad 0")!
861 local($magic_rdata_seen) = 0;
863 # HPPAs, MIPSen: also start medding at chunk 1
865 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
866 $c = $chk[$i]; # convenience copy
868 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
870 # toss all prologue stuff; HPPA is pretty weird
872 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
877 # be slightly paranoid to make sure there's
878 # nothing surprising in there
879 if ( $c =~ /--- BEGIN ---/ ) {
880 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
882 # remove junk whitespace around the split point
886 if ($TargetPlatform =~ /^i386-/) {
887 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
889 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
893 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
894 # when compiling PrimOps.cmm with -optc-O2:
897 # movl %ecx, 16(%esp)
898 # movl %edx, 20(%esp)
899 # but then the code of the function doesn't assume
900 # anything about the contnets of these stack locations.
901 # I think it's to do with the use of inline functions for
902 # PK_Word64() and friends, where gcc is initialising the
903 # contents of the struct to zero, and failing to optimise
904 # away the initialisation. Let's live dangerously and
905 # discard these initalisations.
907 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
908 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
909 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
910 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
911 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
912 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
914 if ($TargetPlatform =~ /^i386-apple-darwin/) {
916 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
918 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
919 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
920 $p =~ s/^\"?L\d+\$pb\"?:\n//;
922 if ($pcrel_reg eq "bx") {
923 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
924 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
928 } elsif ($TargetPlatform =~ /^x86_64-/) {
929 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
930 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
931 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
933 } elsif ($TargetPlatform =~ /^ia64-/) {
934 $p =~ s/^\t\.prologue .*\n//;
936 # Record the number of local and out registers for register relocation later
937 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, (\d+), (\d+), 0\n//;
941 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
942 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
944 # Ignore save/restore of these registers; they're taken
945 # care of in StgRun()
946 $p =~ s/^\t\.save ar\.lc, r\d+\n//;
947 $p =~ s/^\t\.save pr, r\d+\n//;
948 $p =~ s/^\tmov r\d+ = ar\.lc\n//;
949 $p =~ s/^\tmov r\d+ = pr\n//;
951 # Remove .proc and .body directives
952 $p =~ s/^\t\.proc [a-zA-Z0-9_.]+#\n//;
953 $p =~ s/^\t\.body\n//;
955 # If there's a label, move it to the body
956 if ($p =~ /^[a-zA-Z0-9.]+:\n/) {
961 # Remove floating-point spill instructions.
962 # Only fp registers 2-5 and 16-23 are saved by the runtime.
963 if ($p =~ s/^\tstf\.spill \[r1[4-9]\] = f([2-5]|1[6-9]|2[0-3])(, [0-9]+)?\n//g) {
964 # Being paranoid, only try to remove these if we saw a
966 $p =~ s/^\tmov r1[4-9] = r12\n//;
967 $p =~ s/^\tadds r1[4-9] = -[0-9]+, r12\n//g;
968 $p =~ s/^\t\.save\.f 0x[0-9a-fA-F]\n//g;
969 $p =~ s/^\t\.save\.gf 0x0, 0x[0-9a-fA-F]+\n//g;
972 $p =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
973 $p =~ s/^\t\.(mii|mmi|mfi)\n//g; # bundling is no longer sensible
974 $p =~ s/^\t;;\n//g; # discard stops
975 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
977 # GCC 3.3 saves r1 in the prologue, move this to the body
978 # (Does this register get restored anywhere?)
979 if ($p =~ /^\tmov r\d+ = r1\n/) {
983 } elsif ($TargetPlatform =~ /^m68k-/) {
984 $p =~ s/^\tlink a6,#-?\d.*\n//;
985 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
986 # The above showed up in the asm code,
987 # so I added it here.
988 # I hope it's correct.
990 $p =~ s/^\tmovel d2,sp\@-\n//;
991 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
992 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
993 } elsif ($TargetPlatform =~ /^mips-/) {
994 # the .frame/.mask/.fmask that we use is the same
995 # as that produced by GCC for miniInterpret; this
996 # gives GDB some chance of figuring out what happened
997 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
998 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
999 $p =~ s/^\t\.(mask|fmask).*\n//g;
1000 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
1001 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
1002 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
1003 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
1004 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
1005 $p =~ s/__FRAME__/$FRAME/;
1006 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1008 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
1010 $p =~ s/^\tmflr r0\n//;
1011 $p =~ s/^\tbl saveFP # f\d+\n//;
1012 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
1013 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1014 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
1015 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
1016 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
1017 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
1018 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
1019 $p =~ s/^\tbcl 20,31,\"?L\d+\$pb\"?\n//;
1020 $p =~ s/^\"?L\d+\$pb\"?:\n//;
1021 $p =~ s/^\tmflr r31\n//;
1023 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1024 # under some circumstances, only when generating position dependent code.
1025 # I have no idea why, and I don't think it is necessary, so let's toss it.
1026 $p =~ s/^\tli r\d+,0\n//g;
1027 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
1028 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
1029 $p =~ s/^\tmflr 0\n//;
1030 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1031 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
1032 $p =~ s/^\tstw r0,8\(1\)\n//;
1033 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
1034 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1036 # GCC's "large-model" PIC (-fPIC)
1038 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
1040 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
1041 $p =~ s/^.LCF\d+:\n//;
1042 $p =~ s/^\tmflr 30\n//;
1043 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
1044 $p =~ s/^\tadd 30,0,30\n//;
1046 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1047 # under some circumstances, only when generating position dependent code.
1048 # I have no idea why, and I don't think it is necessary, so let's toss it.
1049 $p =~ s/^\tli \d+,0\n//g;
1050 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
1051 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1052 $p =~ s/^\tmr 31,1\n//;
1053 $p =~ s/^\tmflr 0\n//;
1054 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
1055 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
1056 $p =~ s/^\tstd r0,8\(1\)\n//;
1057 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
1058 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
1060 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
1061 # under some circumstances, only when generating position dependent code.
1062 # I have no idea why, and I don't think it is necessary, so let's toss it.
1063 $p =~ s/^\tli \d+,0\n//g;
1064 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
1066 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
1069 # HWL HACK: dont die, just print a warning
1070 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
1071 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
1073 # For PIC, we want to keep part of the prologue
1074 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
1075 # Darwin: load the current instruction pointer into register r31
1076 $p .= "bcl 20,31,$pcrel_label\n";
1077 $p .= "$pcrel_label:\n";
1078 $p .= "\tmflr r31\n";
1079 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
1080 # Linux: load the GOT pointer into register 30
1081 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
1082 $p .= ".LCF$pcrel_label:\n";
1083 $p .= "\tmflr 30\n";
1084 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
1085 $p .= "\tadd 30,0,30\n";
1086 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
1087 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
1088 $p .= "$pcrel_label:\n";
1091 # glue together what's left
1096 if ( $TargetPlatform =~ /^mips-/ ) {
1097 # MIPS: first, this basic sequence may occur "--- END ---" or not
1098 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1101 # toss all epilogue stuff; again, paranoidly
1102 if ( $c =~ /--- END ---/ ) {
1103 # Gcc may decide to replicate the function epilogue. We want
1104 # to process all epilogues, so we split the function and then
1106 @fragments = split(/--- END ---/, $c);
1107 $r = shift(@fragments);
1109 # Rebuild `c'; processed fragments will be appended to `c'
1112 foreach $e (@fragments) {
1113 # etail holds code that is after the epilogue in the assembly-code
1114 # layout and should not be filtered as part of the epilogue.
1116 if ($TargetPlatform =~ /^i386-/) {
1118 $e =~ s/^\tpopl\s+\%edi\n//;
1119 $e =~ s/^\tpopl\s+\%esi\n//;
1120 $e =~ s/^\tpopl\s+\%edx\n//;
1121 $e =~ s/^\tpopl\s+\%ecx\n//;
1122 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1123 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1124 } elsif ($TargetPlatform =~ /^ia64-/) {
1125 # The epilogue is first split into:
1126 # $e, the epilogue code (up to the return instruction)
1127 # $etail, non-epilogue code (after the return instruction)
1128 # The return instruction is stripped in the process.
1129 if (!(($e, $etail) = split(/^\tbr\.ret\.sptk\.many b0\n/, $e))) {
1130 die "Epilogue doesn't seem to have one return instruction: $e\n";
1132 # Remove 'endp' directive from the tail
1133 $etail =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1135 # If a return value is saved here, discard it
1136 $e =~ s/^\tmov r8 = r14\n//;
1138 # Remove floating-point fill instructions.
1139 # Only fp registers 2-5 and 16-23 are saved by the runtime.
1140 if ($e =~ s/^\tldf\.fill f([2-5]|1[6-9]|2[0-3]) = \[r1[4-9]\](, [0-9]+)?\n//g) {
1141 # Being paranoid, only try to remove this if we saw a fill
1143 $e =~ s/^\tadds r1[4-9] = [0-9]+, r12//g;
1146 $e =~ s/^\tnop(?:\.[mifb])?\s+\d+\n//g; # remove nop instructions
1147 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1148 $e =~ s/^\tmov ar\.lc = r\d+\n//;
1149 $e =~ s/^\tmov pr = r\d+, -1\n//;
1150 $e =~ s/^\tmov b0 = r\d+\n//;
1151 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1152 #$e =~ s/^\tbr\.ret\.sptk\.many b0\n//; # already removed
1153 $e =~ s/^\t\.(mii|mmi|mfi|mib)\n//g; # bundling is no longer sensible
1154 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1155 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1156 } elsif ($TargetPlatform =~ /^m68k-/) {
1157 $e =~ s/^\tunlk a6\n//;
1159 } elsif ($TargetPlatform =~ /^mips-/) {
1160 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1161 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1162 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1163 $e =~ s/^\tj\t\$31\n//;
1164 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1165 $e =~ s/^\taddi r1,r1,\d+\n//;
1166 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1167 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1168 $e =~ s/^\tmtlr r0\n//;
1170 $e =~ s/^\tb restFP ;.*\n//;
1171 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1172 $e =~ s/^\tmr 3,0\n//;
1173 $e =~ s/^\taddi 1,1,\d+\n//;
1174 $e =~ s/^\tld 0,16\(1\)\n//;
1175 $e =~ s/^\tmtlr 0\n//;
1177 # callee-save registers
1178 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1179 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1181 # get rid of the debug junk along with the blr
1182 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1184 # incase we missed it with the last one get the blr alone
1187 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1190 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1192 # glue together what's left
1195 $c =~ s/\n\t\n/\n/; # junk blank line
1198 if ($TargetPlatform =~ /^ia64-/) {
1199 # On IA64, remove an .endp directive even if no epilogue was found.
1200 # Code optimizations may have removed the "--- END ---" token.
1201 $c =~ s/^\t\.endp [a-zA-Z0-9_.]+#\n//;
1205 # On SPARCs, we don't do --- BEGIN/END ---, we just
1206 # toss the register-windowing save/restore/ret* instructions
1207 # directly unless they've been generated by function definitions in header
1209 if ( $TargetPlatform =~ /^sparc-/ ) {
1210 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1211 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1213 # throw away PROLOGUE comments
1214 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1217 # On Alphas, the prologue mangling is done a little later (below)
1219 # toss all calls to __DISCARD__
1220 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1221 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1222 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1223 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1224 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1225 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1227 # IA64: fix register allocation; mangle tailcalls into jumps
1228 if ($TargetPlatform =~ /^ia64-/) {
1229 ia64_rename_registers($ia64_locnum, $ia64_outnum) if (defined($ia64_locnum));
1230 ia64_mangle_tailcalls();
1233 # MIPS: that may leave some gratuitous asm macros around
1234 # (no harm done; but we get rid of them to be tidier)
1235 $c =~ s/^\t\.set\tnoreorder\n\t\.set\tnomacro\n\taddu\t(\S+)\n\t\.set\tmacro\n\t\.set\treorder\n/\taddu\t$1\n/
1236 if $TargetPlatform =~ /^mips-/;
1238 # toss stack adjustment after DoSparks
1239 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1240 if $TargetPlatform =~ /^m68k-/; # this looks old...
1242 if ( $TargetPlatform =~ /^alpha-/ &&
1243 ! $magic_rdata_seen &&
1244 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1245 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1246 $magic_rdata_seen = 1;
1249 # pick some end-things and move them to the next chunk
1251 # pin a funny end-thing on (for easier matching):
1252 $c .= 'FUNNY#END#THING';
1254 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1258 # on x86 we try not to copy any directives into a literal
1259 # chunk, rather we keep looking for the next real chunk. This
1260 # is because we get things like
1262 # .globl blah_closure
1268 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1270 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1273 if ( $j < $numchks ) {
1274 $chk[$j] = $to_move . $chk[$j];
1278 elsif ( ( $i < ($numchks - 1)
1279 && ( $to_move =~ /${T_COPY_DIRVS}/
1280 || ( $TargetPlatform =~ /^hppa/
1281 && $to_move =~ /align/
1282 && $chkcat[$i+1] eq 'literal')
1285 || ($to_move =~ /^[ \t]*\.section[ \t]+\.note\.GNU-stack,/)
1287 $chk[$i + 1] = $to_move . $chk[$i + 1];
1288 # otherwise they're tossed
1291 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1294 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1296 # toss all prologue stuff, except for loading gp, and the ..ng address
1297 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1298 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1299 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1300 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1301 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1302 $c = $keep . "..ng:\n";
1304 print STDERR "malformed code block ($ent)?\n"
1307 $c .= "\t.prologue" . $r;
1311 $c =~ s/FUNNY#END#THING//;
1313 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1315 $chk[$i] = $c; # update w/ convenience copy
1318 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1319 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1322 if ( $TargetPlatform =~ /^alpha-/ ) {
1323 # print out the header stuff first
1324 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1325 print OUTASM $chk[0];
1327 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1328 print OUTASM $chk[0];
1330 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1331 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1333 # get rid of horrible "<dollar>Revision: .*$" strings
1334 local(@lines0) = split(/\n/, $chk[0]);
1336 while ( $z <= $#lines0 ) {
1337 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1340 while ( $z <= $#lines0 ) {
1342 last if $lines0[$z] =~ /[,\t]0x0$/;
1348 $chk[0] = join("\n", @lines0);
1349 $chk[0] =~ s/\n\n+/\n/;
1350 print OUTASM $chk[0];
1353 # print out all the literal strings next
1354 for ($i = 0; $i < $numchks; $i++) {
1355 if ( $chkcat[$i] eq 'literal' ) {
1357 # HACK: try to detect 16-byte constants and align them
1358 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1359 # aligned constants, and so does Darwin/x86.
1360 if ( $TargetPlatform =~ /^x86_64/
1361 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1363 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1364 print OUTASM $T_HDR_literal16;
1366 print OUTASM $T_HDR_literal;
1369 print OUTASM $T_HDR_literal;
1372 print OUTASM $chk[$i];
1373 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1375 $chkcat[$i] = 'DONE ALREADY';
1379 # on the HPPA, print out all the bss next
1380 if ( $TargetPlatform =~ /^hppa/ ) {
1381 for ($i = 1; $i < $numchks; $i++) {
1382 if ( $chkcat[$i] eq 'bss' ) {
1383 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1384 print OUTASM $chk[$i];
1386 $chkcat[$i] = 'DONE ALREADY';
1391 # $numchks + 1 as we have the extra one for .note.GNU-stack
1392 for ($i = $FIRST_MANGLABLE; $i < $numchks + 1; $i++) {
1393 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1395 next if $chkcat[$i] eq 'DONE ALREADY';
1397 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1398 if ($chk[$i] ne '') {
1399 print OUTASM $T_HDR_misc;
1400 &print_doctored($chk[$i], 0);
1403 } elsif ( $chkcat[$i] eq 'verbatim' ) {
1404 print OUTASM $chk[$i];
1406 } elsif ( $chkcat[$i] eq 'toss' ) {
1407 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1409 } elsif ( $chkcat[$i] eq 'data' ) {
1410 if ($chk[$i] ne '') {
1411 print OUTASM $T_HDR_data;
1412 print OUTASM $chk[$i];
1415 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1416 # we can just re-constitute this one...
1417 # NB: we emit _three_ underscores no matter what,
1418 # so ghc-split doesn't have to care.
1419 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1421 } elsif ( $chkcat[$i] eq 'closure'
1422 || $chkcat[$i] eq 'srt'
1423 || $chkcat[$i] eq 'infotbl'
1424 || $chkcat[$i] eq 'entry') { # do them in that order
1425 $symb = $chksymb[$i];
1428 if ( defined($closurechk{$symb}) ) {
1429 print OUTASM $T_HDR_closure;
1430 print OUTASM $chk[$closurechk{$symb}];
1431 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1435 if ( defined($srtchk{$symb}) ) {
1436 print OUTASM $T_HDR_relrodata;
1437 print OUTASM $chk[$srtchk{$symb}];
1438 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1442 if ( defined($infochk{$symb}) ) {
1444 print OUTASM $T_HDR_info;
1445 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1447 # entry code will be put here!
1449 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1453 if ( defined($entrychk{$symb}) ) {
1455 $c = $chk[$entrychk{$symb}];
1457 # If this is an entry point with an info table,
1458 # eliminate the entry symbol and all directives involving it.
1459 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1461 foreach $l (split(/\n/,$c)) {
1462 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1464 # If we have .type/.size direrctives involving foo_entry,
1465 # then make them refer to foo_info instead. The information
1466 # in these directives is used by the cachegrind annotator,
1467 # so it is worthwhile keeping.
1468 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1469 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1473 next if $l =~ /^\s*\..*$symb.*\n?/;
1476 $c = join("\n",@o) . "\n";
1479 print OUTASM $T_HDR_entry;
1481 &print_doctored($c, 1); # NB: the 1!!!
1483 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1486 } elsif ( $chkcat[$i] eq 'vector' ) {
1487 $symb = $chksymb[$i];
1490 if ( defined($vectorchk{$symb}) ) {
1491 print OUTASM $T_HDR_vector;
1492 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1494 # direct return code will be put here!
1495 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1497 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1498 # Alphas: the commented nop is for the splitter, to ensure
1499 # that no module ends with a label as the very last
1500 # thing. (The linker will adjust the label to point
1501 # to the first code word of the next module linked in,
1502 # even if alignment constraints cause the label to move!)
1504 print OUTASM "\t# nop\n";
1507 } elsif ( $chkcat[$i] eq 'rodata' ) {
1508 print OUTASM $T_HDR_rodata;
1509 print OUTASM $chk[$i];
1510 $chkcat[$i] = 'DONE ALREADY';
1511 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1512 print OUTASM $T_HDR_relrodata;
1513 print OUTASM $chk[$i];
1514 $chkcat[$i] = 'DONE ALREADY';
1515 } elsif ( $chkcat[$i] eq 'toc' ) {
1516 # silly optimisation to print tocs, since they come in groups...
1517 print OUTASM $T_HDR_toc;
1519 while ($chkcat[$j] eq 'toc')
1520 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1523 print OUTASM $chk[$j];
1525 $chkcat[$j] = 'DONE ALREADY';
1529 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1530 # apple-darwin: dynamic linker stubs
1531 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1532 { # print them out unchanged, but remove the stubs for __DISCARD__
1533 print OUTASM $chk[$i];
1535 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1536 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1537 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1538 print OUTASM $chk[$i];
1540 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1544 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1547 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1548 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1552 On IA64, tail calls are converted to branches at this point. The mangler
1553 searches for function calls immediately followed by a '--- TAILCALL ---'
1554 token. Since the compiler can put various combinations of labels, bundling
1555 directives, nop instructions, stops, and a move of the return value
1556 between the branch and the tail call, proper matching of the tail call
1557 gets a little hairy. This subroutine does the mangling.
1559 Here is an example of a tail call before mangling:
1562 br.call.sptk.many b0 = b6
1577 sub ia64_mangle_tailcalls {
1578 # Function input and output are in $c
1580 # Construct the tailcall-mangling expression the first time this function
1582 if (!defined($IA64_MATCH_TAILCALL)) {
1583 # One-line pattern matching constructs. None of these
1584 # should bind references; all parenthesized terms
1585 # should be (?:) terms.
1586 my $stop = q/(?:\t;;\n)/;
1587 my $bundle = q/(?:\t\.(?:mii|mib|mmi|mmb|mfi|mfb|mbb|bbb)\n)/;
1588 my $nop = q/(?:\tnop(?:\.[mifb])?\s+\d+\n)/;
1589 my $movgp = q/(?:\tmov r1 = r\d+\n)/;
1590 my $postbr = q/(?:\tbr \.L\d+\n)/;
1592 my $noeffect = "(?:$stop$bundle?|$nop)*";
1593 my $postbundle = "(?:$bundle?$nop?$nop?$postbr)?";
1595 # Important parts of the pattern match. The branch target
1596 # and subsequent jump label are bound to $1 and $2
1597 # respectively. Sometimes there is no label.
1598 my $callbr = q/^\tbr\.call\.sptk\.many b0 = (.*)\n/;
1599 my $label = q/(?:^\.L([0-9]*):\n)/;
1600 my $tailcall = q/\t--- TAILCALL ---\n/;
1602 $IA64_MATCH_TAILCALL =
1603 $callbr . $label . '?' . $noeffect . $movgp . '?' . $noeffect .
1604 $tailcall . $stop . '?' . '(?:' . $postbundle . ')?';
1607 # Find and mangle tailcalls
1608 while ($c =~ s/$IA64_MATCH_TAILCALL/\tbr\.few $1\n/o) {
1609 # Eek, the gcc optimiser is getting smarter... if we see a jump to the
1610 # --- TAILCALL --- marker then we reapply the substitution at the source sites
1611 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1614 # Verify that all instances of TAILCALL were processed
1615 if ($c =~ /^\t--- TAILCALL ---\n/) {
1616 die "Unmangled TAILCALL tokens remain after mangling"
1621 The number of registers allocated on the IA64 register stack is set
1622 upon entry to the runtime with an `alloc' instruction at the entry
1623 point of \verb+StgRun()+. Gcc uses its own `alloc' to allocate
1624 however many registers it likes in each function. When we discard
1625 gcc's alloc, we have to reconcile its register assignment with what
1628 There are three stack areas: fixed registers, input/local registers,
1629 and output registers. We move the output registers to the output
1630 register space and leave the other registers where they are.
1633 sub ia64_rename_registers() {
1634 # The text to be mangled is in $c
1635 # Find number of registers in each stack area
1636 my ($loc, $out) = @_;
1642 # These are the register numbers used in the STG runtime
1643 my $STG_FIRST_OUT_REG = 32 + 34;
1644 my $STG_LAST_OUT_REG = $STG_FIRST_OUT_REG + 7;
1646 $first_out_reg = 32 + $loc;
1648 if ($first_out_reg > $STG_FIRST_OUT_REG) {
1649 die "Too many local registers allocated by gcc";
1652 # Split the string into fragments containing one register name each.
1653 # Rename the register in each fragment and concatenate.
1655 foreach $fragment (split(/(?=r\d+[^a-zA-Z0-9_.])/s, $c)) {
1656 if ($fragment =~ /^r(\d+)((?:[^a-zA-Z0-9_.].*)?)$/s) {
1659 if ($regnum < $first_out_reg) {
1660 # This is a local or fixed register
1662 # Local registers 32 and 33 (r64 and r65) are
1663 # used to hold saved state; they shouldn't be touched
1664 if ($regnum == 64 || $regnum == 65) {
1665 die "Reserved register $regnum is in use";
1669 # This is an output register
1670 $regnum = $regnum - $first_out_reg + $STG_FIRST_OUT_REG;
1671 if ($regnum > $STG_LAST_OUT_REG) {
1672 die "Register number ($regnum) is out of expected range";
1676 # Update this fragment
1677 $fragment = "r" . $regnum . $2;
1688 sub hppa_mash_prologue { # OK, epilogue, too
1691 # toss all prologue stuff
1692 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1694 # Lie about our .CALLINFO
1695 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1702 # toss all epilogue stuff
1703 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1705 # Sorry; we moved the _info stuff to the code segment.
1706 s/_info,DATA/_info,CODE/g;
1713 sub print_doctored {
1714 local($_, $need_fallthru_patch) = @_;
1716 if ( $TargetPlatform =~ /^x86_64-/ ) {
1719 # movq -4(%ebp), %rax
1724 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1725 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1728 if ( $TargetPlatform !~ /^i386-/
1729 || ! /^\t[a-z]/ # no instructions in here, apparently
1730 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1735 # OK, must do some x86 **HACKING**
1737 local($entry_patch) = '';
1738 local($exit_patch) = '';
1740 # gotta watch out for weird instructions that
1741 # invisibly smash various regs:
1742 # rep* %ecx used for counting
1743 # scas* %edi used for destination index
1744 # cmps* %e[sd]i used for indices
1745 # loop* %ecx used for counting
1750 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1752 # * GCC used an "STG reg" for its own purposes
1754 # * some secret uses of machine reg, requiring STG reg
1755 # to be saved/restored
1757 # The most dangerous "GCC uses" of an "STG reg" are when
1758 # the reg holds the target of a jmp -- it's tricky to
1759 # insert the patch-up code before we get to the target!
1760 # So here we change the jmps:
1762 # --------------------------------------------------------
1763 # it can happen that we have jumps of the form...
1764 # jmp *<something involving %esp>
1766 # jmp <something involving another naughty register...>
1768 # a reasonably-common case is:
1770 # movl $_blah,<bad-reg>
1773 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1777 # movl -4(%ebx), %eax
1782 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1784 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1785 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1786 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1787 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1788 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1789 if /(jmp|call)\s+.*\%esi/;
1791 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1792 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1793 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1794 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1795 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1796 if /(jmp|call)\s+.*\%edi/;
1799 # OK, now we can decide what our patch-up code is going to
1802 # Offsets into register table - you'd better update these magic
1803 # numbers should you change its contents!
1804 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1805 local($OFFSET_Hp)=88;
1807 # Note funky ".=" stuff; we're *adding* to these _patch guys
1808 if ( $StolenX86Regs <= 2
1809 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1810 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1811 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1813 # nothing for call_{entry,exit} because %esi is callee-save
1815 if ( $StolenX86Regs <= 3
1816 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1817 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1818 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1820 # nothing for call_{entry,exit} because %edi is callee-save
1823 # --------------------------------------------------------
1824 # next, here we go with non-%esp patching!
1826 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1828 # Before calling GC we must set up the exit condition before the call
1829 # and entry condition when we come back
1831 # fix _all_ non-local jumps:
1833 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1834 # On Darwin, we've got local-looking jumps that are
1835 # actually global (i.e. jumps to Lfoo$stub or via
1836 # Lfoo$non_lazy_ptr), so we fix those first.
1837 # In fact, we just fix everything that contains a dollar
1838 # because false positives don't hurt here.
1840 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1843 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1844 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1846 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1848 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1849 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1851 if ($StolenX86Regs == 2 ) {
1852 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1853 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1854 } elsif ($StolenX86Regs == 3 ) {
1855 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1856 if /^\t(jmp|call)\s+.*\%edi/;
1859 # --------------------------------------------------------
1860 # that's it -- print it
1862 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1866 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1867 print OUTASM $exit_patch;
1868 # ToDo: make it not print if there is a "jmp" at the end
1874 sub init_FUNNY_THINGS {
1875 %KNOWN_FUNNY_THING = (
1877 # "${T_US}stg_.*{T_POST_LBL}", 1,
1882 The following table reversal is used for both info tables and return
1883 vectors. In both cases, we remove the first entry from the table,
1884 reverse the table, put the label at the end, and paste some code
1885 (that which is normally referred to by the first entry in the table)
1886 right after the table itself. (The code pasting is done elsewhere.)
1890 local($symb, $tbl, $discard1) = @_;
1892 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1894 local($before) = '';
1896 local(@imports) = (); # hppa only
1899 local(@lines) = split(/\n/, $tbl);
1902 # Deal with the header...
1903 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1904 $label .= $lines[$i] . "\n",
1905 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1906 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1907 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1909 $before .= $lines[$i] . "\n"; # otherwise...
1913 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1915 # Grab the table data...
1916 if ( $TargetPlatform !~ /^hppa/ ) {
1917 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1919 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1920 # to offsets (relative to the info label),
1921 # in order to support position independent code.
1922 $line =~ s/$infoname/0/
1923 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1924 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1925 || $line =~ s/([A-Za-z0-9_]+_str)$/\1 - $infoname/
1926 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1927 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1928 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1929 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1930 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1931 push(@words, $line);
1933 } else { # hppa weirdness
1934 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1935 # FIXME: the RTS now expects offsets instead of addresses
1936 # for all labels in info tables.
1937 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1938 push(@imports, $lines[$i]);
1940 # We don't use HP's ``function pointers''
1941 # We just use labels in code space, like normal people
1942 $lines[$i] =~ s/P%//;
1943 push(@words, $lines[$i]);
1948 # Now throw away any initial zero word from the table. This is a hack
1949 # that lets us reduce the size of info tables when the SRT field is not
1950 # needed: see comments StgFunInfoTable in InfoTables.h.
1952 # The .zero business is for Linux/ELF.
1953 # The .skip business is for Sparc/Solaris/ELF.
1954 # The .blockz business is for HPPA.
1956 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1961 for (; $i <= $#lines; $i++) {
1962 $after .= $lines[$i] . "\n";
1965 # Alphas: If we have anonymous text (not part of a procedure), the
1966 # linker may complain about missing exception information. Bleh.
1967 # To suppress this, we place a .ent/.end pair around the code.
1968 # At the same time, we have to be careful and not enclose any leading
1969 # .file/.loc directives.
1970 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1971 local ($ident) = $1;
1972 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1973 $after .= "\t.end $ident\n";
1976 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1977 # assembler (!) wherein .quad constants inside .text sections are
1978 # first narrowed to 32 bits then sign-extended back to 64 bits.
1979 # This obviously screws up our 64-bit bitmaps, so we work around
1980 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1981 if ( $TargetPlatform =~ /^alpha-/ ) {
1983 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1984 local ($number) = $1;
1985 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1986 local ($sign, $base, $digits) = ($1, $2, $3);
1987 $base = (10, 8, 16)[length $base];
1988 local ($hi, $lo) = (0, 0);
1989 foreach $i (split(//, $digits)) {
1990 $j = $lo * $base + $i;
1991 $lo = $j % 4294967296;
1992 $hi = $hi * $base + ($j - $lo) / 4294967296;
1994 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1996 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1997 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1999 print STDERR "Cannot handle \".quad $number\" in info table\n";
2006 if ( $TargetPlatform =~ /x86_64-apple-darwin/ ) {
2007 # Tack a label to the front of the info table, too.
2008 # For now, this just serves to work around a crash in Apple's new
2009 # 64-bit linker (it seems to assume that there is no data before the
2010 # first label in a section).
2012 # The plan for the future is to do this on all Darwin platforms, and
2013 # to add a reference to this label after the entry code, just as the
2014 # NCG does, so we can enable dead-code-stripping in the linker without
2015 # losing our info tables. (Hence the name _dsp, for dead-strip preventer)
2017 $before .= "\n${infoname}_dsp:\n";
2021 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
2022 . join("\n", @words) . "\n"
2025 # print STDERR "before=$before\n";
2026 # print STDERR "label=$label\n";
2027 # print STDERR "words=",(reverse @words),"\n";
2028 # print STDERR "after=$after\n";
2034 The HP is a major nuisance. The threaded code mangler moved info
2035 tables from data space to code space, but unthreaded code in the RTS
2036 still has references to info tables in data space. Since the HP
2037 linker is very precise about where symbols live, we need to patch the
2038 references in the unthreaded RTS as well.
2041 sub mini_mangle_asm_hppa {
2042 local($in_asmf, $out_asmf) = @_;
2044 open(INASM, "< $in_asmf")
2045 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
2046 open(OUTASM,"> $out_asmf")
2047 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
2050 s/_info,DATA/_info,CODE/; # Move _info references to code space
2056 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
2057 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
2063 sub tidy_up_and_die {
2064 local($return_val, $msg) = @_;
2066 exit (($return_val == 0) ? 0 : 1);